Pilgrim, kadischi, and stateless (was Re: [Fedora-livecd-list] pilgrim livecd work)

[Toshio Kuratomi]

On Thu, 2006-09-21 at 20:31 -0400, David Zeuthen wrote:
> Hey Dave,
> 
> On Wed, 2006-09-20 at 20:20 -0400, Dave Jones wrote:
> > On Wed, Sep 20, 2006 at 04:07:12PM -0400, David Zeuthen wrote:
> > 
> >  > Btw, If someone could talk davej into including unionfs into the Fedora
> >  > kernel, we'd use that instead of dm-snapshot and we'd have persistence
> >  > more easily solved [1].
> > 
> > I think the comments Al Viro had on it the last time it was reviewed
> > were for the most part unprintable. I wouldn't hold your breath for
> > this to appear any time soon.
> 
> Right. My understanding is that the controversial part of unionfs is the
> ability to join multiple writable file systems into a single tree. Is
> this correct?
> 
> If so, note that this is not a feature livecd nor stateless needs, the
> one part is always read-only, the other parts is just a single overlay
> where we take damage.
> 
> How hard would it be to do a unionfs-ro (read only) with the following
> semantics
> 
>  1. Support exactly two underlying directories, the first assumed to be
>     read only
> 
>  2. If some file exists in both trees, pick file with latest ctime
> 
> I dunno much about the kernel VFS layer to say whether this is easy or
> hard but I do hope this is a lot simpler than what current unionfs is
> doing. So.. is this hard?
> 
> Btw, justification for 2. ("pick file with latest ctime", not just "if
> file exist in rw branch, always pick rw branch"): suppose I use a livecd
> using this unionfs-ro fs and updates my bash package. The changes are
> now written to a USB stick such that I have a persistent session. I now
> download a newer version of the livecd where the bash package is newer.
> When using this together with my USB stick, we'll pick the
> newest /bin/bash file.
> 
This sounds interesting but problematic.  You can't free the space at
this point.  Using touch could cause a file to disappear with no way to
recover it.  Etc, etc.

Multiple writable fs's should be okay as long as the merged filesystem
only writes to a single fs.  Say the last writable fs added to the
stack.  Allowing writes to several layers in the stack is... not
something I want to attempt :-)  Maybe you could create a simple unionfs
using fuse, although it might sacrifice performance for convenience....

-Toshio
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: This is a digitally signed message part
Url : http://lists.fedoraproject.org/pipermail/livecd/attachments/20060921/5982f0dc/attachment.bin